| license: apache-2.0 | |
| task_categories: | |
| - unconditional-image-generation | |
| tags: | |
| - diffusion | |
| - classifier-free-guidance | |
| - tiny-model | |
| - pytorch | |
| # PocketDiffusion | |
| PocketDiffusion is a compact class-conditional denoising diffusion model for 8x8 | |
| handwritten digits. It learns to predict Gaussian noise over 50 diffusion steps and | |
| uses classifier-free guidance during sampling. | |
| The same frozen Tiny Vision classifier used for GlyphForge evaluates conditional | |
| recognizability, making the VAE and diffusion results directly comparable under one | |
| judge. | |
| ## Reproduce | |
| ```powershell | |
| uv run python projects/tiny-vision-foundry/prepare_data.py | |
| uv run python projects/pocket-diffusion/train.py | |
| ``` | |
| ## Verified results | |
| - Parameters: **55,608** | |
| - Diffusion steps: **50** | |
| - Training epochs: **300** | |
| - Generated samples: **1,000** | |
| - Selected classifier-free guidance: **3.0** | |
| - Frozen-judge class fidelity: **96.20%** | |
| Guidance search improved fidelity monotonically from 46.10% at `1.0` to 96.20% at | |
| `3.0`. Per-class fidelity ranged from 83% for digit `8` to 100% for digits `0` and | |
| `6`. Mean within-class pixel variance ranged from 0.0209 to 0.0456, noticeably higher | |
| than the CVAE's 0.0066 to 0.0168 range under the same 100-samples-per-class protocol. | |